Enhancing image inpainting through image decomposition and deep neural networks

Abstract

A new approach to inpainting problems that combines domain decomposition methods (DDM) with deep neural networks (DNN) to solve partial differential equations (PDE) is presented. First, this article examines different existing and emerging approaches to inpainting while emphasizing their advantages and disadvantages in a unified framework. After that, we introduce an algorithm that highlights the combination of DDM and DNN techniques for solving PDEs of a proposed mathematical inpainting model. For this model, the modified approach that has been adopted uses the DNN method which is based on convolutional neural networks (CNN) to reduce the computational cost in our algorithm while maintaining accuracy. Finally, the experimental results show that our method significantly outperforms existing ones for high-resolution images in paint stains.